R help - Mar 2009 - intelligent optimizer (with domain restrictions?)

dear R experts---sorry, second question of the day. I want to match some  
moments. I am writing my own code---I have exactly as many moment  
conditions as parameters, and I am leary of having to learn the magic of  
GMM weighting matrices (if I was to introduce more). the process sounds  
easy conceptually. (Seen it in seminars many times, so how hard could it  
possibly be?...me thinks) first time I am trying this. some of my moments  
are standard deviations. Easy, me thinks. Just maximize the  
exp(my.sigma.parameter) instead of the my.sigma.parameter. This way, nlm()  
can throw negative values into my objective function, and I will be good.  
this is about the time to start laughing, of course.

so, nlm() computes a gradient that is huge at my initial starting value. it  
then decides that it wants to take a step into exp(20.59), at which point  
everything in my function goes heywire and it wants to return NA. now nlm()  
barfs...and I am seriously consider grid-searching. This does not strike me  
as particular intelligent.

are there any intelligent optimizers that understand domains and/or  
will "backstep" gracefully when they encounter an NA? are there better
ways
to deal with matching second moments?

advice appreciated.

regards,

/iaw

PS: you probably don't want to know this, but I have a dynamic panel data  
set; and my goal is to test whether a constant auto-coefficient across  
units can describe the data. that is, I want to find out whether x(i,t)= a  
+ b(i) + c*x(i,t-1) is better replaced by x(i,t)=a + b(i) + c(i)*x(i,t-1).  
right now, I am running N OLS TS regression of x on lagged x, and am  
picking off the mean(c), sd(c), and mean(sigma_i) and sd(sigma_i). if there  
is a procedure in R that already does a test for heterogeneous  
autocorrelation coefficients in a more intelligent fashion, please please  
point me to it. however, even if this exists, I think I need to figure out  
how to find a more graceful optimizer anyway.

	[[alternative HTML version deleted]]

Hi,

Without knowing much about the problem, it is difficult to provide good advice. 
Having said that, it seems like you are trying to solve a system of nonlinear
equations by matching theoretical moments to their empirical counterparts.   You
can do this by using a nonlinear equations solver such as dfsane() in the the
package "BB" or nleqslv() in "nleqslv".

It is not clear to me how you end up with a scalar objective function to
minimize (do you consider the L2-norm of the residuals?).

Ravi.

____________________________________________________________________

Ravi Varadhan, Ph.D.
Assistant Professor,
Division of Geriatric Medicine and Gerontology
School of Medicine
Johns Hopkins University

Ph. (410) 502-2619
email: rvaradhan at jhmi.edu


----- Original Message -----
From: ivowel at gmail.com
Date: Wednesday, March 25, 2009 6:16 pm
Subject: [R] intelligent optimizer (with domain restrictions?)
To: r-help <r-help at stat.math.ethz.ch>

> dear R experts---sorry, second question of the day. I want to match 
> some  
>  moments. I am writing my own code---I have exactly as many moment  
>  conditions as parameters, and I am leary of having to learn the magic 
> of  
>  GMM weighting matrices (if I was to introduce more). the process 
> sounds  
>  easy conceptually. (Seen it in seminars many times, so how hard could 
> it  
>  possibly be?...me thinks) first time I am trying this. some of my 
> moments  
>  are standard deviations. Easy, me thinks. Just maximize the  
>  exp(my.sigma.parameter) instead of the my.sigma.parameter. This way, 
> nlm()  
>  can throw negative values into my objective function, and I will be 
> good.  
>  this is about the time to start laughing, of course.
>  
>  so, nlm() computes a gradient that is huge at my initial starting 
> value. it  
>  then decides that it wants to take a step into exp(20.59), at which 
> point  
>  everything in my function goes heywire and it wants to return NA. now 
> nlm()  
>  barfs...and I am seriously consider grid-searching. This does not 
> strike me  
>  as particular intelligent.
>  
>  are there any intelligent optimizers that understand domains and/or  
> 
>  will "backstep" gracefully when they encounter an NA? are there 
> better ways  
>  to deal with matching second moments?
>  
>  advice appreciated.
>  
>  regards,
>  
>  /iaw
>  
>  PS: you probably don't want to know this, but I have a dynamic panel 
> data  
>  set; and my goal is to test whether a constant auto-coefficient 
> across  
>  units can describe the data. that is, I want to find out whether 
> x(i,t)= a  
>  + b(i) + c*x(i,t-1) is better replaced by x(i,t)=a + b(i) + 
> c(i)*x(i,t-1).  
>  right now, I am running N OLS TS regression of x on lagged x, and am  
> 
>  picking off the mean(c), sd(c), and mean(sigma_i) and sd(sigma_i). if 
> there  
>  is a procedure in R that already does a test for heterogeneous  
>  autocorrelation coefficients in a more intelligent fashion, please 
> please  
>  point me to it. however, even if this exists, I think I need to 
> figure out  
>  how to find a more graceful optimizer anyway.
>  
>  	[[alternative HTML version deleted]]
>  
>  ______________________________________________
>  R-help at r-project.org mailing list
>  
>  PLEASE do read the posting guide 
>  and provide commented, minimal, self-contained, reproducible code.